Perturbing BEAMs: EEG adversarial attack to deep learning models for epilepsy diagnosing

Table 8 Transferability of the adversarial samples when the source model is BEAM-related model and the target model is EEG-related models. The N of GPBEAM-DE is 5

Architecture			Maxpool		Temporal convolution		LSTM		Mixed LSTM
\({\mathrm{DL}}_{E}\)	\(\epsilon\)	Evaluation Criteria	Acc	SR	Acc	SR	Acc	SR	Acc	SR
0	-	-	0.92	-	0.84	-	0.90	-	0.88	-
0.024	0.1	\(\mathrm{GPBEAM}\)	0.92	0.01	0.84	0.01	0.90	0.01	0.88	0.01
0.021	0.1	GPBEAM-DE	0.92	0.01	0.84	0.01	0.90	0.01	0.88	0.01
0.030	0.3	\(\mathrm{GPBEAM}\)	0.92	0.01	0.84	0.02	0.90	0.01	0.88	0.01
0.026	0.3	GPBEAM-DE	0.92	0.01	0.84	0.01	0.90	0.01	0.88	0.01
0.036	0.5	\(\mathrm{GPBEAM}\)	0.91	0.02	0.84	0.03	0.90	0.01	0.88	0.02
0.030	0.5	GPBEAM-DE	0.92	0.01	0.84	0.01	0.90	0.01	0.88	0.01

ISSN: 1472-6947